【论文学习一】A fusion approach to classify hyperspectral oil spill data 2018

目录

 

## Summary

## Research Objective

## Problem Statement

## Method(s)

## Evaluation

## Conclusion

strong conclusion:

weak conclusion：

## Notes

---

## Summary

溢油特征仅存在于光谱特征的某些波段中，因此，在不关注正确波段的情况下选择特征可能会导致选择对目标识别不太重要的特征。

## Research Objective

Study and extract the signal characteristics that emphasize spectral information in the context of oil spill. One way to extract the features is to apply the filter banks on a spectral region that emphasizes oil sensitive reflectance zones.

研究和提取在石油泄漏情况下强调光谱信息的信号特性。提取要素的一种方法是在强调油敏反射区的光谱区域上应用滤波器组

 

## Problem Statement

the oil features are present only in certain bands of a spectral signature. Hence, feature selection without focusing on right bands may lead to selection of features which are less significant for target identification.

溢油特征仅存在于光谱特征的某些波段中，因此，在不关注正确波段的情况下选择特征可能会导致选择对目标识别不太重要的特征。

## Method(s)

 

1. In this study oil feature were selected by applying filter bank to the modeled oil emulsion spectral signal. Further additional features were extracted through wavelet transform（小波变换） to take advantage of different but complimentary benefits.
2. The obtained features are fused together and fed to Gaussian Mixture Model (GMM) classifier to classify oil emulsions.

1. 本研究通过将滤波器组应用于模拟油乳化光谱信号，选取了油特征。通过小波变换提取了其他附加功能，以利用不同但适用的好处。公式如下：

       

 

 

 

1. 所得特征融合在一起，并馈送到高斯混合模型（GMM）分类器，对油乳液进行分类。

 

## Evaluation

1. To evaluate the performance of the proposed method, comparison was done with conventional feature extraction methods PCA [6] and Wavelet. In signal modeling various scales such as First Order Derivative (FOD), Second Order Derivative (SOD), Normal Scale (NS) and Mel Scale were used.A detailed description of the Mel Frequency Cepstral Coefficient (MFCC) has been givenin .Mel Scale is given by Eq.
2. To ensure robustness, samples were selected randomly for testing .The experiment was tested for 30, 60, 70 and 100 training samples. Class-specific accuracy, Overall Accuracy (OA), and Kappa coefficient (κ) are used to evaluate the classification accuracy results.

1. 为了评价该方法的性能，对传统的特征提取方法PCA[6]和小波进行了比较。在信号建模中，使用了一阶导数（FOD）、二阶导数（SOD）、正态量程（NS）和梅尔比例等各种尺度。
2. 为了确保鲁棒性，随机选择样本进行测试。该实验对30、60、70和100个训练样本进行了测试。某类特定精度、总体精度 （OA） 和 Kappa 系数用于评估分类精度结果。

## Conclusion

strong conclusion:

1. It is worth noting that due to the close similarity between different classes, coefficients extracted by applying only one technique (feature selection or feature extraction) would not be enough to discriminate between the classes. However, since the proposed method takes advantage of feature selection through signal modeling and feature extraction through wavelet transform, fused features are able to classify targets with higher accuracy.
2. An oil emulsion signal is broken down into constituent sinusoids of different frequencies through Fourier transform so while rectangular filter banks are placed, some characteristics like drift, trends and abrupt changes may be lost. These characteristics are often the most important part of the signal, however wavelets have the ability to perform local analysis and reveal aspects of data such as self-similarity, trends and discontinuities in higher derivatives. As the proposed method benefits from both signal modeling and wavelet transformation, thus facilitating higher classification accuracy.

1. 由于不同类之间的相似性，仅应用一种技术（特征选择或特征提取）提取的系数不足以区分类。但是，由于本文所提出的方法通过信号建模和通过小波变换提取要素来利用特征选择，熔融特征能够以更高的精度对目标进行分类。
2. 通过傅立叶变换，油乳液信号被分解成不同频率的正弦，因此，当放置矩形滤芯时，一些混光、趋势和突变等字符的成形物可能会丢失。这些特征通常是信号最重要的部分，但是小波能够执行局部分析，并揭示数据的各个方面，如高导数中的自相似性、趋势和不连续性。由于该方法同时受益于信号建模和小波变换，因此有助于提高分类精度.

weak conclusion：

1. The current study focused only on spectral features while spatial features were not considered. Future research could lay emphasis on adding spatial features to improve robustness of the technique for automatic target classification.

1. 目前的研究只关注光谱特征，而不考虑空间特征。未来的研究可以强调增加空间特征，以提高自动目标分类技术的鲁棒性。

 

## Notes

1. 从文献中可以明显看出高光谱成像（HSI）用于石油泄漏检测和分类。与ISODATA、最小距离、决策树等传统技术相比，采用高光谱特征匹配方法[42]，效果更好[32]。
2. [32]Salem F, Kafatos M (2001) Hyperspectral image analysis for oil spill mitigation. In: Paper presented at the 22nd Asian conference on remote sensing vol. 5, p 9 [42]Yuhas RH, Goetz AF, Boardman JW (1992) Discrimination among semi-arid landscape endmembers using the spectral angle mapper (SAM) algorithm. In Proc. Summaries 3rd Annu. JPL Airborne Geosci Workshop(pp 147–149)
3. [2]对五种不同的溢油检测技术进行了比较分析，认为通过频谱滤波可以改善溢油检测。
4. [2] Alam MS, Sidike P (2012) Trends in oil spill detection via hyperspectral imaging. In: Electrical & ComputerEngineering (ICECE), 2012 7th International Conference on. pp 858–862

1. 虽然HSI为准确的目标分类提供了足够的光谱信息，但由于训练样品需求增加，高光谱维数会影响精度[14]。
2. [14]Donoho DL (2000) High-dimensional data analysis: the curses and blessings of dimensionality. AMS Math Challenges Lecture 1(2000):32